Skip to content

Which version of a page a visitor gets

One precedence decides every arm: a matched audience wins, then a locked winner, then the split assignment, then Default.

One function decides which arm of a variant container a visitor sees, whether the container personalizes a whole section or a single module inside one. The precedence lives in exactly one place, so a personalized band and a personalized module on the same page can never disagree about the same visitor.

The order

Four checks run, in this order, for every container:

  1. A matched audience arm always wins, whichever mode the container is in.
  2. In an A/B test, a locked winner routes everyone the split would have bucketed to that arm — including visitors who would otherwise have bucketed somewhere else.
  3. Still in an A/B test, the visitor's own server-resolved split assignment for that experiment.
  4. The Default arm, when nothing above applies.

One precedence, not two that happen to agree

Audiences are checked once, ahead of the split, inside the one selection that every container goes through. Nothing on the experiment side ever sees an audience; it only works out which split arm this visitor belongs to, and that answer is consulted after the audience check.

The same answer also drives what is physically in the page. Before the document crosses to the browser, the server strips the arms this visitor wasn't served, using the selection that just ran — so what gets pruned and what gets rendered cannot disagree. What that leaves visible is What is public about a personalized page.

The deterministic split

When nothing above has decided it, assignment is a pure hash of the experiment id and the visitor id — not a lookup against anything stored. The same visitor lands on the same arm on every later request without a row being written first; only whether that exposure gets recorded is a separate decision. The mechanics of the hash and how a split gets configured belong to Running an A/B experiment.

What happens when none of the above applies at all — no audience, no locked winner, no experiment even running — is its own page: The default arm is the real page.